Enhance response API support to not fail with tool calling

Author: luis5tb

docs/static/llama-stack-spec.html

@@ -6170,25 +6170,13 @@
             "OpenAIResponseInput": {
                 "oneOf": [
                     {
-                        "$ref": "#/components/schemas/OpenAIResponseOutputMessageWebSearchToolCall"
-                    },
-                    {
-                        "$ref": "#/components/schemas/OpenAIResponseOutputMessageFileSearchToolCall"
-                    },
-                    {
-                        "$ref": "#/components/schemas/OpenAIResponseOutputMessageFunctionToolCall"
+                        "$ref": "#/components/schemas/OpenAIResponseOutput"
                     },
                     {
                         "$ref": "#/components/schemas/OpenAIResponseInputFunctionToolCallOutput"
                     },
-                    {
-                        "$ref": "#/components/schemas/OpenAIResponseMCPApprovalRequest"
-                    },
                     {
                         "$ref": "#/components/schemas/OpenAIResponseMCPApprovalResponse"
-                    },
-                    {
-                        "$ref": "#/components/schemas/OpenAIResponseMessage"
                     }
                 ]
             },

docs/static/llama-stack-spec.yaml

@@ -4588,13 +4588,9 @@ components:
         Error details for failed OpenAI response requests.
     OpenAIResponseInput:
       oneOf:
-        - $ref: '#/components/schemas/OpenAIResponseOutputMessageWebSearchToolCall'
-        - $ref: '#/components/schemas/OpenAIResponseOutputMessageFileSearchToolCall'
-        - $ref: '#/components/schemas/OpenAIResponseOutputMessageFunctionToolCall'
+        - $ref: '#/components/schemas/OpenAIResponseOutput'
         - $ref: '#/components/schemas/OpenAIResponseInputFunctionToolCallOutput'
-        - $ref: '#/components/schemas/OpenAIResponseMCPApprovalRequest'
         - $ref: '#/components/schemas/OpenAIResponseMCPApprovalResponse'
-        - $ref: '#/components/schemas/OpenAIResponseMessage'
     "OpenAIResponseInputFunctionToolCallOutput":
       type: object
       properties:

llama_stack/apis/agents/openai_responses.py

@@ -746,15 +746,7 @@ class OpenAIResponseInputFunctionToolCallOutput(BaseModel):
 
 OpenAIResponseInput = Annotated[
     # Responses API allows output messages to be passed in as input
-    OpenAIResponseOutputMessageWebSearchToolCall
-    | OpenAIResponseOutputMessageFileSearchToolCall
-    | OpenAIResponseOutputMessageFunctionToolCall
-    | OpenAIResponseInputFunctionToolCallOutput
-    | OpenAIResponseMCPApprovalRequest
-    | OpenAIResponseMCPApprovalResponse
-    |
-    # Fallback to the generic message type as a last resort
-    OpenAIResponseMessage,
+    OpenAIResponseOutput | OpenAIResponseInputFunctionToolCallOutput | OpenAIResponseMCPApprovalResponse,
     Field(union_mode="left_to_right"),
 ]
 register_schema(OpenAIResponseInput, name="OpenAIResponseInput")

llama_stack/providers/inline/agents/meta_reference/responses/streaming.py

@@ -125,7 +125,10 @@ async def create_response(self) -> AsyncIterator[OpenAIResponseObjectStream]:
         while True:
             # Text is the default response format for chat completion so don't need to pass it
             # (some providers don't support non-empty response_format when tools are present)
-            response_format = None if self.ctx.response_format.type == "text" else self.ctx.response_format
+
+            response_format = (
+                None if getattr(self.ctx.response_format, "type", None) == "text" else self.ctx.response_format
+            )
             completion_result = await self.inference_api.openai_chat_completion(
                 model=self.ctx.model,
                 messages=messages,

llama_stack/providers/inline/agents/meta_reference/responses/tool_executor.py

@@ -371,7 +371,8 @@ async def _build_result_messages(
                     content.append(part)
             else:
                 raise ValueError(f"Unknown result content type: {type(result.content)}")
-            input_message = OpenAIToolMessageParam(content=content, tool_call_id=tool_call_id)
+            # OpenAI tool messages must have simple string content, not complex content structures
+            input_message = OpenAIToolMessageParam(content=str(content), tool_call_id=tool_call_id)
         else:
             text = str(error_exc) if error_exc else "Tool execution failed"
             input_message = OpenAIToolMessageParam(content=text, tool_call_id=tool_call_id)

llama_stack/providers/utils/inference/openai_compat.py

@@ -861,10 +861,11 @@ def _convert_openai_request_response_format(
 
 
 def _convert_openai_tool_calls(
-    tool_calls: list[OpenAIChatCompletionMessageFunctionToolCall],
+    tool_calls: list[OpenAIChatCompletionMessageFunctionToolCall] | list[ToolCall],
 ) -> list[ToolCall]:
     """
     Convert an OpenAI ChatCompletionMessageToolCall list into a list of ToolCall.
+    If already ToolCall objects, return them as-is.
 
     OpenAI ChatCompletionMessageToolCall:
         id: str
@@ -885,12 +886,15 @@ def _convert_openai_tool_calls(
     if not tool_calls:
         return []  # CompletionMessage tool_calls is not optional
 
+    # Convert from OpenAI format to ToolCall format if not already ToolCall objects
     return [
         ToolCall(
             call_id=call.id,
             tool_name=call.function.name,
             arguments=call.function.arguments,
         )
+        if not isinstance(call, ToolCall)
+        else call
         for call in tool_calls
     ]
 
@@ -955,13 +959,19 @@ def _convert_openai_sampling_params(
 
 
 def openai_messages_to_messages(
-    messages: list[OpenAIMessageParam],
+    messages: list[OpenAIMessageParam] | list[Message],
 ) -> list[Message]:
     """
     Convert a list of OpenAIChatCompletionMessage into a list of Message.
+    If already Message objects, return them as-is.
     """
     converted_messages = []
     for message in messages:
+        # Check if this individual message is already a Llama Stack Message
+        if isinstance(message, (UserMessage | SystemMessage | ToolResponseMessage | CompletionMessage)):
+            # Already a Llama Stack Message, use as-is
+            converted_messages.append(message)
+            continue
         if message.role == "system":
             converted_message = SystemMessage(content=openai_content_to_content(message.content))
         elif message.role == "user":
@@ -973,9 +983,11 @@ def openai_messages_to_messages(
                 stop_reason=StopReason.end_of_turn,
             )
         elif message.role == "tool":
+            # Handle both OpenAI format (tool_call_id) and Llama Stack format (call_id)
+            tool_call_id = getattr(message, "tool_call_id", None) or getattr(message, "call_id", None)
             converted_message = ToolResponseMessage(
                 role="tool",
-                call_id=message.tool_call_id,
+                call_id=tool_call_id,
                 content=openai_content_to_content(message.content),
             )
         else: